Kernel: stable AF_UNIX UAF fix, kbuild speedups, THP and futex work

Stable maintainers pressed for an AF_UNIX use-after-free fix on long-term trees while mainline discussion focused on faster builds and several memory-management series. Futex, NUMA, compaction, and BPF patches also advanced with concrete performance and locking questions.

Stable AF_UNIX use-after-free fix for older trees

A patch aimed at 6.1.y repairs a peek-then-tail dereference use-after-free in unix_stream_data_wait(), tracked as CVE-2026-64109. The stable maintainer is seeking an AF_UNIX acknowledgment for a custom lock approach that would also cover 5.15 and 5.10. Operators of those long-term kernels face a security exposure until the fix lands.

kbuild series targets large build speedups

Version 3 of a twenty-patch kbuild series claims up to 36 percent faster allmodconfig builds. Linus Torvalds and other maintainers are examining the implementation details on linux-kernel. Frequent rebuilders have a direct stake in whether the claimed gains survive review.

Always-enable mTHP support

A fourteen-patch v8 series refactors has_transparent_hugepage() usage and adds pgtable_has_pmd_leaves() so multi-size transparent huge pages can be enabled on architectures that lack PMD-sized pages. The work broadens THP availability beyond traditional PMD assumptions. Non-x86 memory-management users stand to gain more flexible huge-page coverage.

RFC for stealable FUTEX_PING locks

An RFC proposes FUTEX_PING, a stealable futex that uses proxy execution to deliver priority-inheritance locks that can be stolen. Peter Zijlstra and others raised deadlock and locking-order concerns. Real-time and lock-heavy workloads would be affected if the design moves forward.

PMD-level swap entries for anonymous THPs

A resend of the twenty-nine-patch v7 set introduces PMD-level swap entries so anonymous transparent huge pages can survive swap without splitting to PTEs. Discussion continues among multiple memory-management developers. Preserving THP integrity across swap would cut fragmentation costs for large anonymous mappings.

NUMA promotion no longer gated by VMA scan filters

A four-patch v2 series stops VMA scan filters from blocking NUMA promotion, letting read-only file mappings and shared folios move from slow tiers. Promotion scans are decoupled from placement scans. Tiered-memory systems could see more effective automatic page migration.

Direct compaction skipped for costly GFP_NORETRY

A v4 page_alloc patch avoids direct compaction for costly __GFP_NORETRY allocations, recovering a 2.6x throughput drop seen in fragmented PostgreSQL workloads after an iomap change. Multiple developers reviewed the approach. Allocation-heavy and database workloads on fragmented machines are the clearest beneficiaries.

BPF user-memory kfuncs land in bpf-next

Version 9 of a five-patch series that adds bpf_copy_from_user_mm kfuncs and trusted-or-null marking for bprm->mm was merged into bpf-next. The helpers let BPF programs holding an mm_struct perform safer user-memory access. BPF authors gain new primitives for process-memory inspection.